GOOGLE / GOOGLE DEEPMIND · MODEL RELEASE TRACKER
Gemini 4 Argon
Gemini 4 Argon is a Gemini 4 frontier model from Google described as optimized for deep, long‑horizon reasoning and for software‑engineering and knowledge‑work workflows. It was released prior to the announcement of Gemini 4 Carbon and is referenced in Google’s Gemini 4 family materials.CURRENT SNAPSHOT5/5 DIMENSIONS WITH DATA
The dimensions that change the decision.
RELEASE TIMELINE
Published, source-backed release events only.
Google introduces Gemini Robotics 2.5, a more capable vision-language-action model for real-world dexterity
Google announced Gemini Robotics 2.5, an upgraded vision-language-action (VLA) model that increases task breadth, dexterity, reliability, and safety for robots. Google says the model was evaluated on over 200 tasks (showing 30–100% higher success rates and 20–30% faster completion versus prior baselines) and achieved a 96% success rate in a Mercedes‑Benz proof‑of‑concept using Apptronik’s Apollo 2 humanoid on production lines.
Google releases Gemini 4 Argon for robotics control, simulation, and policy improvement
Google released Gemini 4 Argon, a general-purpose physical-reasoning model aimed at robot control, Real2Sim scene construction, multimodal labeling of robot video, and iterative policy improvement. Google’s demos show Argon guiding a Boston Dynamics Spot and robot arms in MuJoCo twins, reconstructing scenes from teleoperation video, and synthesizing domain-randomized MuJoCo episodes (reported to raise success from 56% to 75% on one setup); Google reports Argon scores 87% on the ERQA benchmark, and makes it available to Google AI Ultra subscribers and via the paid Gemini API as gemini-4-argon, while publishing only limited per-demo failure data and relying on heavy prompt scaffolding in examples.
Google launches Gemini 4 Argon, a new frontier model, globally
Google today made Gemini 4 Argon globally available to developers and enterprises via the API in AI Studio and the Gemini platform, and to Google AI Ultra/Pro subscribers in Antigravity, the Gemini app, and AI Mode in Search. Google positions Argon as a high‑precision frontier model for complex multi‑step tasks across coding, finance, legal, cybersecurity, and science, citing top ranking on the Vals Index, example deployments inside Google (Chrome refactor, ADK optimizations), technical speedups (xTB and Gemma inference via vLLM), and introductory pricing ($2 per million input tokens, $10 per million output tokens, with cached inputs 95% off).
Google releases Nano Banana 2.1 image model and halves generation prices
Google released Nano Banana 2.1, an image-generation and editing model running on Gemini 3.6 Flash that the company says improves visual quality, text rendering, character consistency, wide panoramas, and infographics while supporting up to 14 reference images and keeping up to four characters and ten objects consistent. Google also cut Nano Banana pricing roughly in half (for example, 1K images from 6.70¢ to 3.36¢), is rolling 2.1 out across products including the Gemini app, AI Mode in Google Search, Google AI Studio, Flow and Stitch, and will shut down the prior gemini-3.1-flash-image model on October 29, 2026.
Google DeepMind unveils Gemini 4 Argon, a frontier model with 1M-token context
Google DeepMind announced Gemini 4 Argon, a new frontier multimodal model optimized for long-horizon reasoning and complex workflows. Argon is rolling out initially to trusted cyber defenders via the Fairwind Program, expands context length to 1 million tokens, reports leading benchmark performance across coding, finance, legal, and video understanding, and will be made more widely available after phased safety testing and engagement with U.S. government pre-release processes; Google also published introductory pricing.
VERIFIABLE FACTS
Every value stays attached to a source and date.
Google describes Gemini 4 Argon as having been launched recently (it preceded the announcement of Gemini 4 Carbon).
Available to Google AI Ultra subscribers and via the paid Gemini API (model id 'gemini-4-argon' in the Gemini API docs).
Google describes Argon as built to sustain deep reasoning across complex, long‑horizon workflows and delivering frontier performance in real‑world software engineering, enterprise knowledge work (legal, finance), and cybersecurity defense.
Reported introductory price: $2 per million input tokens and $10 per million output tokens (stated as the same introductory price as Gemini 4 Argon).
Google reports Argon helped optimize spacetime resources (qubits × gates) for quantum subroutines and in one example beat the published baseline by 40% in minutes.
Google says Argon agents analyzed fleet‑wide profiling telemetry to identify and apply memory optimizations that freed over 300 TiB once rolled out, with an estimated 500 TiB to 1 PiB in total savings.
Google reports Argon agents are being used for large‑scale codebase migrations and optimizations (e.g., migrating C/C++ codebases to Rust across Google, scaling up to 800K+ lines for the Fuchsia Zircon kernel).
Early practitioner reaction highlighted concerns that simulation twins can look visually convincing while object/contact dynamics behave incorrectly, which matters for manipulation performance.
Argon is described as built with an industry‑leading one‑million‑token context capability (the reporting also states support for up to one million output tokens).
Argon is engineered for advanced/deep reasoning across complex, long‑horizon workflows and is positioned to excel at heavy‑duty workloads such as coding, cybersecurity defense, enterprise knowledge work (legal, finance), and autonomous cybersecurity patching.
Reporting states Argon was trained and described as able to autonomously find, validate, and patch critical software vulnerabilities (used particularly for defensive cyber work).
Media reporting cites Google claims that Argon scored higher than competing frontier models (including OpenAI’s GPT‑6 Astra and Anthropic’s Fable/Opus) on a variety of AI benchmarks and is listed as leading on at least one model index cited by Google.
Media reporting states that the next major gain in image quality 'could' come with Gemini 4 Argon — phrasing indicates a potential improvement but not a confirmed capability or benchmark in the provided excerpts.
Described as a multimodal physical-reasoning model combining visual inputs, language, and action/simulation for robotics (video + language + control).
Reported as likely to provide the "next major gain" in image quality compared with recent image models (e.g., Gemini 3.6 Flash / Nano Banana 2.1); presented as a possibility rather than a confirmed benchmark result.
Google presented Gemini 4 Argon as its newest 'frontier' model and said it was 'rolled out Argon broadly last week' in a blog post published Oct 23, 2026, indicating a broad rollout in mid–late October 2026.
Google describes Argon as designed for deep work in complex, real‑world tasks across coding, finance, legal, cybersecurity, and science, with improved multi‑step reasoning, reduced hedging and repetition, and higher precision on long tasks.
Reported score of 87% on the Embodied Reasoning Question Answering (ERQA) benchmark using pure reasoning without helper tools; reported by Google as leading other frontier models on that test.
Google reports Argon ranks #1 on the Vals Index, an internal measure mentioned as assessing model performance across economic sectors such as finance, coding, legal, and tax work.
Google reports Argon solved all 12 end‑to‑end simulated cyberattack scenarios in their cybersecurity tests (up from five in the previous generation), and describes frontier‑level cyberdefense capabilities intended to help defenders.
On DeepMind's RE‑Bench (74 realistic ML engineering tasks), Gemini 4 Argon solves about half of the tasks (~50%). By comparison, Gemini 3.7 Flash solved 27% on the same benchmark.
Across 1,389 internal Argon agent runs, the success rate without human input dropped sharply as task complexity increased, indicating reduced ability to fully automate more complex tasks.
DeepMind reports Gemini 4 Argon sits slightly above the Gemini capability trend line; the team attributes this position to increased compute and organizational investment rather than AI‑driven acceleration of research.
Google describes Argon as excelling at deep, multi-step reasoning and structural precision across complex, real-world tasks in coding, finance, legal, cybersecurity, and science; examples include large codebase refactoring, code optimization, finding and patching cybersecurity vulnerabilities, and supporting scientific research.
Independent reporting and the public demos cited by Google indicate Argon’s examples still rely on detailed human prompting and additional agentic verification steps; demos show multi-step human involvement and verification before physical execution.
Google says Argon 'excels at deep work in complex, real‑world tasks' across multiple domains including coding, finance, legal, cybersecurity, and science.
Argon scored 87% on the Embodied Reasoning (ERQA) benchmark, reported to outperform other frontier models on that benchmark.
In reporting by testers, Argon scored 97.5% on ARC-AGI-3 when run with a "Provider Adapter Harness" (a test setup that uses vendor-specific API features).
Under the ARC-AGI "Standard Harness" (which allows only limited notes between rounds), testers reported Argon dropping to 49.8% at low reasoning; testers also reported API issues that prevented completing all games in that harness.
Described as excelling at deep, multi-step tasks across domains such as coding, finance, legal, cybersecurity, and science; emphasized strengths include physical understanding, code generation, and visual‑semantic reasoning for robotics and real‑world workflows.
Reported 97.5% on ARC‑AGI‑3 when tested with a vendor‑specific setup ("Provider Adapter Harness"); the report notes the result depends heavily on the testing method and environment.
In the ARC standard harness (which discards internal reasoning between actions), Argon’s reported performance drops to 49.8% at low reasoning; testers also reported API issues that prevented completing all games in the standard harness.
Reportedly used by Wiz to find a critical PHP vulnerability; described as especially useful for finding the vulnerability, developing exploits, building attack detection, and validating the patch.
Google states that Argon's depth and precision across multi‑step tasks leads it to outperform other frontier models on benchmarks across domains.
The blog reports 'major upgrades in spatial reasoning, mathematical precision, and visual coding,' enabling Argon to generate complex 3D environments and coordinate motions of multiple elements.
An 11‑hour unattended run of Gemini 4 Argon produced two accepted optimizations for Google Agent Development Kit's codebase.
Argon read repository guidelines, established performance baselines, proposed optimizations, reviewed its own changes, and ran all 16,892 open‑source unit tests for the repository.
Argon operated on a codebase spanning over 200,000 lines of Python across approximately 800 files.
The accepted optimizations reduced in‑memory session‑preparation overhead by about 40% in long sessions.
The accepted optimizations reduced cold package import time by 21%.
Google states Gemini 4 Argon can reason, write code, and coordinate tools to execute complex, long‑horizon workflows and deep comprehension tasks.
In an example, Gemini 4 Argon (in Antigravity) parsed more than 80,000 raw files from NASA — including weather logs and engineering, health, and science data across 72 flights on Mars — and organized insights into an executive‑ready deck.
Google shows Argon producing 5,000 lines of math to render a 3D sunlit library from scratch and then creating a video about the process, demonstrating combined mathematical and 3D‑generation abilities.
A research experiment reported that a Gemini 4 Argon model broke down tasks into steps that the Genie 4 world model then simulated in real time, illustrating Argon’s use for task decomposition and planning in simulation/robotics contexts.
Argon scored 50 on the Artificial Analysis Intelligence Index v5.0 (reported as Argon's score, with Carbon higher at 59).
Reported to perform well in areas such as finance, legal, cybersecurity, and science; cited use cases include untangling legacy codebases, accelerating scientific research, robotics, and other domain-specific tasks.
Positioned for complex coding, financial analysis, and legal workflows.
Framed as a general-purpose physical-reasoning model for robot control, Real2Sim scene construction, multimodal labeling of robot video, and iterative policy improvement rather than as a single end-effector policy.
Google reports synthesizing domain-randomized MuJoCo episodes from one RGB photo to fine-tune a vision-language-action policy, raising success on the corresponding real setup from 56% to 75%.
Google's robotics examples use long, task-specific prompts, coordinate grids, motion primitives, and verifiers (scaffolding is part of the system under test); Google does not publish per-demo failure rates beyond the single VLA case.
Presented as a general-purpose physical-reasoning model for robot control, Real2Sim scene construction (video-to-sim), multimodal labeling of robot video, vision-language-action policies, and iterative policy improvement. Examples in the report include directing a Boston Dynamics Spot, steering an arm with visual coordinates and HTTP motion primitives, and closing a loop with the Gemini Omni video world model.
Reported demos and evaluations rely on long, task-specific prompts, coordinate grids, motion primitives and other scaffolding as part of the system; the article notes Google does not publish per-demo failure rates beyond a single VLA case.
Framed as a general-purpose physical-reasoning model for robot control, Real2Sim scene construction, multimodal labeling of robot video, and iterative policy improvement; reported uses include steering simulation, synthesizing domain-randomized episodes, and improving real-world policy success rates.
Argon scores 87% on the Embodied Reasoning Question Answering (ERQA) benchmark using pure reasoning without helper tools.
The Argon robotics post did not publish per-demo failure rates beyond a single VLA case (Google did not publish per-demo failure rates beyond that VLA example).
Argon is described as a frontier model that "excels at deep, long‑horizon reasoning" and as the ideal choice for the majority of software‑engineering and knowledge‑work tasks.
On Google’s cited Artificial Analysis Intelligence Index at the default “high” thinking level, Argon is reported to score 50 (compared with Hydrogen’s 51; Hydrogen reaches 54 at its highest “xhigh” setting, per Google).
Reporting indicates Argon offers only medium and high thinking levels (contrasting with Hydrogen’s low, medium, high, and new “xhigh” options).
WHAT CHANGED